For us non-Folding DCer's, does the GPU support work?

[GoHack]

For us non-Folding DCer's, does GPU support really indeed work as advertised on the Folding@Home?

If it does work, I'd like to see it utilized on other DC projects.

I wish other DC projects put as much effort in utilizing our computer hardware as Folding@Home does.

Thx

 

[Hito Bahadur]

It works. It's nascent, but it does seem to produce results. It would do even better if I quit tweaking.

 

[Pocatello]

It works. It's nascent, but it does seem to produce results. It would do even better if I quit tweaking.

I have a suggestion, senor Bahadur. Lay off the caffeine... and the tweaking will go away.

 

[Bill Clo]

It looks like it'll get better. Beta software, beta drivers, totally new concept that's never been tried before. Some people are with the program and are being helpful, some are already bitching about the "points". Give it some time to mature.

I suspect that some of the other DC projects will at least think about using GPUs now that FAH has shown the way. Ask your favorite project's developers about it if you can.

 

[GoHack]

Is there a big increase in number crunching, or just a minor one?

Not being familiar w/Folding@Home,on how long it takes to do one Work Unit, if for example it takes 4 hours to do 1 Work Unit without the GPU, how long would it take to run that same Work Unit with the GPU support?

 

[Bill Clo]

Is there a big increase in number crunching, or just a minor one?

Not being familiar w/Folding@Home,on how long it takes to do one Work Unit, if for example it takes 4 hours to do 1 Work Unit without the GPU, how long would it take to run that same Work Unit with the GPU support?

From what Stanford says, the GPU does the number crunching about 20-50x faster than a regular CPU. If you go here, and look at the Teraflops of calculations, you see that about 328 GPUs are doing more than the work of 16000 (!) Linux based machines. http://fah-web.stanford.edu/cgi-bin/main.py?qtype=osstats

The work units are totally different; there is a GPU-only sub-type.

 

[mhouston]

The main trick is getting the algorithm to match the computational model of a streaming processor. This is somewhat similar to how you code for Cell, at least in my experience thus far, so the effort isn't wasted. (There are other coding styles that are more thread based, but streaming approaches seem to work pretty well as you can use async DMAs to move data...) Not all algorithms are a good match. We really need something that requires large amounts of compute per offchip memory read. The folding guys do LOTS of math per read with the current algorithms running on the GPU, so this was relatively easy once we found the right algorithm match. [Vishal who did most of the heavy lifting with the final code might disagree ;-) ]. However, they have other algorithms that we haven't figured out how to get running nearly as well, but we are taking another look based on what we've learned.

If other DC projects want to take a look at doing this, they can contact the Brook team at Stanford, or others in the GPGPU space (http://www.gpgpu.org), and we can try to help you. FAH GPU was all done in Brook.